I simply extracted the model weights from their checkpoints, so they are loadable in PyTorch.
You can verify that the files contain model weights that are identical to the checkpoints.
The verification process will be slow, because you need to download 33GB of the original checkpoints.
1python -m venv .venv
2source .venv/bin/activate
3pip install huggingface-hub torch torchvision tqdm
4python verify.py
1import torch
2from torchvision.models import resnet18
3
4model = resnet18()
5model.load_state_dict(torch.load("resnet18_l2_eps1.pth", weights_only=True))